Re: [Lse-tech] Re: CPU affinity & IPI latency

[Shailabh Nagar]

David,

> Global scheduling decisions should be triggered in response of load
unbalancing
> and not at each schedule() run otherwise we're going to introduce a
common lock
> that will limit the overall scalability.

Thats correct. Though it beggars the question : who will define
"load-imbalance" and at what granularity ? In the Loadbalancing extensions
to MQ (http://lse.sourceforge.net/scheduling/LB/poolMQ.html) load balancing
is done at a frequency specified at the time the loadbalancing module is
loaded. The parameter can be changed dynamically through a /proc interface.
So we are providing a knob for the user/sysadmin to control the
loadbalancing desired.

> My idea about the future of the scheduler is to have a config options
users can
> chose depending upon the machine use.
> By trying to keep a unique scheduler for both UP and MP is like going to
give
> the same answer to different problems and the scaling factor ( of the
scheduler
> itself ) on SMP will never improve.

That is true to an extent. It would be convenient for us as scheduler
rewriters to have neatly differentiated classes like UP, SMP, BIG_SMP, NUMA
etc. But it forces all other scheduler-sensitive code to think of each of
these cases separately and is exactly the reason why #ifdef's are
discouraged for critical kernel code like the scheduler.

Its certainly a challenge to provide SMP/NUMA scalability in the scheduler
(and elsewhere in the kernel) without having to resort to an #ifdef.

Shailabh